Real-time digital holographic microscopy observable in multi-view and multi-resolution

We propose a real-time digital holographic microscopy, that enables simultaneous multiple reconstructed images with arbitrary resolution, depth and positions, using Shifted-Fresnel diffraction instead of Fresnel diffraction. In this system, we used four graphics processing units (GPU) for multiple reconstructions in real-time. We show the demonstration of four reconstruction images from a hologram with arbitrary depths, positions, and resolutions.

💡 Research Summary

The paper presents a real‑time digital holographic microscopy (DHM) system capable of simultaneously generating multiple reconstructed images with arbitrary depths, lateral positions, and spatial resolutions. Traditional DHM relies on Fresnel diffraction for reconstruction, which ties the sampling interval and propagation distance to a fixed grid; consequently, obtaining images at different depths or resolutions typically requires multiple recordings or computationally intensive interpolation. To overcome these constraints, the authors adopt the Shifted‑Fresnel diffraction (SFD) formulation. SFD introduces a controllable lateral shift of the input wavefront and allows independent scaling of the sampling grid, while still being implementable with fast Fourier transform (FFT) operations. This retains the O(N log N) computational complexity of conventional Fresnel‑based methods but adds the flexibility to change viewing depth, field‑of‑view, and pixel size on the fly.

The hardware architecture consists of a high‑resolution image sensor that captures a single hologram (≈1920 × 1080 pixels, ~1 GB per frame) and a four‑GPU cluster that performs parallel reconstructions. Each GPU is assigned a distinct set of SFD parameters (propagation distance, sampling interval, shift) and processes the hologram independently. The implementation uses CUDA streams and event‑driven synchronization: the hologram is transferred to each GPU’s global memory, a 2‑D FFT is executed, a phase‑correction kernel (derived from the SFD equations) is applied, and an inverse FFT yields the complex‑amplitude reconstruction. Phase‑correction matrices are pre‑loaded into texture memory to minimize latency. By overlapping data transfer with computation, the overall pipeline achieves sub‑30 ms latency, enabling real‑time display of four views.

Experimental validation demonstrates four reconstructions from a single hologram at depths of 0.5 mm, 1.0 mm, 1.5 mm, and 2.0 mm. The lateral resolutions are independently set to 1 µm/pixel, 2 µm/pixel, 3 µm/pixel, and 5 µm/pixel, respectively. The high‑resolution view reveals fine cellular structures such as microtubules, while the lower‑resolution, larger‑field view provides contextual information. This simultaneous multi‑scale, multi‑depth capability eliminates the need for repeated recordings and reduces experimental variability.

The authors acknowledge several limitations. GPU memory caps the number of concurrent reconstructions and the maximum image size; numerical stability of the complex phase term can degrade for extreme scaling factors; and the current system is limited to four views, dictated by the number of GPUs. Future work proposes scaling to larger GPU clusters or FPGA‑based accelerators, integrating adaptive phase‑correction algorithms to improve numerical robustness, and coupling the platform with machine‑learning‑driven autofocus and resolution‑selection modules. Extending the framework to real‑time 3‑D volume rendering would further broaden its applicability.

In summary, by combining the flexibility of Shifted‑Fresnel diffraction with high‑throughput GPU parallelism, the paper delivers a DHM platform that breaks the traditional trade‑off between depth, resolution, and speed. This advancement opens new possibilities for rapid, multi‑scale optical inspection in biomedical imaging, semiconductor metrology, and optical design verification, where both high spatial fidelity and real‑time feedback are essential.